encrypted fine-tuning.
Tune a model on data the trainer never reads.
Encrypted fine-tuning lets you adapt a model to private data without ever shipping that data — the gradients are computed on ciphertext, so the trainer improves a model it never actually saw.
statusroadmap · q3 2026
dataencrypted at rest + in flight
methodfhe training
purposeprivate adaptation
train on what you can't share
Hospitals with patient records, funds with proprietary signals, firms with privileged documents — all want models that know what they know, without the data leaving their control. That's the point of this feature.